1. EECC722 - Shaaban #1 lec # 6 Fall 2005 9-26-2005 Dynamic Branch Prediction Dynamic branch prediction schemes run-time behavior of branches to make predictions. Usually information about outcomes of previous occurrences of branches are used to predict the outcome of the current branch. One-level or Bimodal: Uses a Branch History Table (BHT), a table of usually two-bit saturating counters which is indexed by a portion of the branch address. Taxonomy of Basic Two-Level Adaptive Branch Predictors. (BP-1) Path-based Correlated Branch Predictors. Hybrid Predictor: Uses a combinations of two or more branch prediction mechanisms. (BP-2) Branch Prediction Aliasing/Intereference. Mechanisms that try to solve the problem of aliasing: –MCFarling’s Two-Level with index sharing (gshare), 1993. (BP-2) –The Agree Predictor, 1997. (BP-5, BP-10) –The Bi-Mode Predictor, 1997. (BP-6, BP-10) –The Skewed Branch Predictor, 1997. (BP-7, BP-10) –YAGS (Yet Another Global Scheme), 1998. (BP-10) BP-10 also covers a good overview of BP 5, 6, 7

2. EECC722 - Shaaban #2 lec # 6 Fall 2005 9-26-2005 One-Level Bimodal Branch Predictors One-level or bimodal branch prediction uses only one level of branch history. These mechanisms usually employ a table which is indexed by lower bits of the branch address. The table entry consists of n history bits, which form an n-bit automaton. Smith proposed such a scheme, known as the Smith algorithm, that uses a table of two-bit saturating counters. One rarely finds the use of more than 3 history bits in the literature. Two variations of this mechanism: –Decode History Table: Consists of directly mapped entries. –Branch History Table (BHT): Stores the branch address as a tag. It is associative and enables one to identify the branch instruction during IF by comparing the address of an instruction with the stored branch addresses in the table. From 551

3. EECC722 - Shaaban #3 lec # 6 Fall 2005 9-26-2005 One-Level Bimodal Branch Predictors Decode History Table (DHT) a Low Bits of Table has 2 a entries. 0 0 1 1 0 1 Not Taken Taken High bit determines branch prediction 0 = Not Taken 1 = Taken Example: For a =12 Table has 2 a = 2 12 entries = 4096 = 4k entries Number of bits needed = 2 x 4k = 8k bits From 551 Or Pattern History Table (PHT)

4. EECC722 - Shaaban #4 lec # 6 Fall 2005 9-26-2005 One-Level Bimodal Branch Predictors Branch History Table (BHT) High bit determines branch prediction 0 = Not Taken 1 = Taken From 551

5. EECC722 - Shaaban #5 lec # 6 Fall 2005 9-26-2005 Basic Dynamic Two-Bit Branch Prediction: Two-bit Predictor State Transition Diagram Transition Diagram From 551 11 10 0100 Or 2-bit saturating counter

6. EECC722 - Shaaban #6 lec # 6 Fall 2005 9-26-2005 Prediction Accuracy of A 4096-Entry Basic Dynamic Two-Bit Branch Predictor Integer average 11% FP average 4% (more loops) Integer From 551

7. EECC722 - Shaaban #7 lec # 6 Fall 2005 9-26-2005 Correlating Branches Recent branches are possibly correlated: The behavior of recently executed branches affects prediction of current branch. Example: Branch B3 is correlated with branches B1, B2. If B1, B2 are both not taken, then B3 will be taken. Using only the behavior of one branch cannot detect this behavior. B1if (aa==2) aa=0; B2if (bb==2) bb=0; B3if (aa!==bb){ SUBIR3, R1, #2 BENZR3, L1 ; b1 (aa!=2) ADDR1, R0, R0 ; aa==0 L1: SUBIR3, R1, #2 BNEZR3, L2 ; b2 (bb!=2) ADDR2, R0, R0 ; bb==0 L2 SUBR3, R1, R2 ; R3=aa-bb BEQZR3, L3 ; b3 (aa==bb) From 551

8. EECC722 - Shaaban #8 lec # 6 Fall 2005 9-26-2005 Two-Level Adaptive Predictors Two-level adaptive predictors were originally proposed by Yeh and Patt (1991). They use two levels of branch history. The first level stored in a Branch History Register (BHR), or Table (BHT), usually one or more k-bit shift register(s). The data in this register is used to index the second level of history, the Pattern History Table (PHT) or tables (PHTs). Yeh and Patt later identified nine variations of this mechanism depending on how branch history and pattern history is kept: per address, globally or per set, plus they give a taxonomy (1993). BP-1

9. EECC722 - Shaaban #9 lec # 6 Fall 2005 9-26-2005 A General Two-level Branch Predictor or Path History Register (PHR) Examples: Per Address First Level: Low”a” bits of branch address (per address first level) Per Set First Level (with s=4 sets): i index bits of branch address used to divide branches into s = 2 i sets e.g In paper #1 (BP-1) i=2 index bits 10, 11 used as branch set index to divide branches into s =4 sets and select between four BHRs in first level Examples: Concatenate low address bits and BHR (GAp) XOR low address bits and BHR (Gshare) Register (BHR) or Table (BHT) First Level: Branch address Set index bits 11 10 Second Level: Or PHTs

10. EECC722 - Shaaban #10 lec # 6 Fall 2005 9-26-2005 Taxonomy of Two-level Adaptive Branch Prediction Mechanisms BP-1 Per Address First Level: “a” low bits of branch address used to select from b = 2 a BHT entries Per Set First Level: i index bits of branch address used to divide branches into s = 2 i sets e.g In paper #1 (BP-1) i=2 index bits 10, 11 used as branch set index to divide branches into s =4 sets and select between four BHRs in first level First LevelSecond Level

11. EECC722 - Shaaban #11 lec # 6 Fall 2005 9-26-2005 Hardware cost of Two-level Adaptive Prediction Mechanisms Hardware cost of Two-level Adaptive Prediction Mechanisms Neglecting logic cost, branch targets and assuming 2-bit of pattern history for each PHT entry. The parameters are as follows: –k is the length of the history registers BHRs, –b is the number of BHT entries, and or number of PHTs in per address schemes b = 2 a where a is the number of low branch address used –s = 2 i (i number of set index bits) is the number of sets of branches in first level (# of BHRs). –p = 2 m (m number of set index bits) is the number of sets of branches in second level (# of PHTs). BP-1

12. EECC722 - Shaaban #12 lec # 6 Fall 2005 9-26-2005 Variations of global history Two-Level Adaptive Branch Prediction. BP-1 One PHT p = 2 m SPHTs b = 2 a PPHTs

13. EECC722 - Shaaban #13 lec # 6 Fall 2005 9-26-2005 Correlating Two-Level Dynamic GAp Branch Predictors Improve branch prediction by looking not only at the history of the branch in question but also at that of other branches: –Record the pattern of the k most recently executed branches as taken or not taken globally in BHR. –Use that pattern to select the proper PHT. In general, the notation: (k, n) GAp predictor means: –Record last k branches to select between 2 k PHTs. –Each table uses n-bit counters (each table entry has n bits). Basic two-bit Bimodal BHT is then a (0,2) predictor. From 551

14. EECC722 - Shaaban #14 lec # 6 Fall 2005 9-26-2005 Organization of A Correlating Two- level GAp (2,2) Branch Predictor First Level (2 bit shift register) Second Level k = # of branches tracked in first level = 2 Thus 2 m = 2 2 = 4 tables in second level a = # of low bits of branch address used = 4 Thus each table in 2nd level has 2 a = 2 4 = 16 entries n = # number of bits of 2nd level table entry = 2 Number of bits for 2nd level = 2 k x n x 2 a = 4 x 2 x 16 = 128 bits High bit determines branch prediction 0 = Not Taken 1 = Taken Low 4 bits of address Selects correct table Selects correct entry in table GAp Global (1st level) Adaptive per address (2nd level) (BHR) (PHTs) From 551

15. EECC722 - Shaaban #15 lec # 6 Fall 2005 9-26-2005 Prediction Accuracy of Two-Bit Dynamic Predictors Under SPEC89 BasicBasicCorrelatingTwo-level GAp From 551

16. EECC722 - Shaaban #16 lec # 6 Fall 2005 9-26-2005 Performance of Global history schemes with different branch history lengths (k = 2-18 bits) The average prediction accuracy of integer (int) and floating point (fp) programs by using global history schemes. These curves are cut of when the implementation cost exceeds 512K bits. BP-1 For 9 SPEC89 benchmarks GAg int GAg fp (GAg) k (GAp) (GAs)

17. EECC722 - Shaaban #17 lec # 6 Fall 2005 9-26-2005 Performance of Global history schemes with different number of pattern history tables (1-1024 PHTs) BP-1 For 9 SPEC89 benchmarks GAg 1 PHT These curves are cut of when the implementation cost exceeds 512K bits.

18. EECC722 - Shaaban #18 lec # 6 Fall 2005 9-26-2005 Most Cost-effective Global: 8K: GAs(7, 32) 128K: GAs (11, 32) BHR length = k (here k= 7 or 11) 32 = 2 5 = Number of PHTs For 9 SPEC89 benchmarks

19. EECC722 - Shaaban #19 lec # 6 Fall 2005 9-26-2005 Notes On Global Schemes Performance The performance of global history schemes is sensitive to branch history register (BHR) length. –Using one PHT prediction accuracy is still rising with 18-bit BHR, improving 25% when BHR is increased from 2 to 18 bits. –Using 16 PHTs, prediction accuracy improved 10% when BHR is increased from 2 to 18 bits. –Lengthening the BHR increases the probability that the history the current branch needs is in the BHR and reduces pattern history (PHT) interference. The performance of global history schemes is also sensitive to the number of pattern history tables (PHTs). –The pattern history of different branches interfere with each other if they map to the same PHT. –Adding more PHTs reduces the probability of pattern history interference by mapping interfering branches to different tables

20. EECC722 - Shaaban #20 lec # 6 Fall 2005 9-26-2005 Variations of per-address history Two-Level Adaptive Branch Prediction BP-1 p = 2 m SPHTs b = 2 a PPHTs Size of BHT = b = 2 a BHRs each k bits Set Index bits

21. EECC722 - Shaaban #21 lec # 6 Fall 2005 9-26-2005 Performance of Per-address history schemes with different branch history lengths (k= 2-18 bits) BP-1 9 SPEC89 benchmarks These curves are cut of when the implementation cost exceeds 512K bits. PAg int PAg fp (PAg) FP Better than Global schemes (BHR)

22. EECC722 - Shaaban #22 lec # 6 Fall 2005 9-26-2005 Performance of Per-address history schemes with different number of pattern history tables (1-1024 PHTs) BP-1 9 SPEC89 benchmarks These curves are cut of when the implementation cost exceeds 512K bits. PAg 1 PHT PHTs FP Better than Global schemes Integer worse than Global schemes k

23. EECC722 - Shaaban #23 lec # 6 Fall 2005 9-26-2005 Most Cost-effective Per address: 8K: PAs(6, 16) 128K: PAs (8, 256) Number of PHTs Here (16 or 256) BHR length = k (here k= 6 or 8) PHTs

24. EECC722 - Shaaban #24 lec # 6 Fall 2005 9-26-2005 The prediction accuracy of per-address history schemes is not as sensitive to BHR length or number of PHTs as in global schemes –Using one PHT, integer prediction accuracy only improves 6% (compared to 25% for global) when BHR is increased from 2 to 18 bits. –Floating point performance is nearly flat when BHR length is larger than 6 bits. –Increasing number of PHTs results in a small performance improvement. Average integer performance of global schemes is higher than per-address schemes: –Due to the large number of if-then-else statements in integer code which depend on the results of adjacent branches, therefore global schemes with better branch correlation perform better. Average floating point performance of per-address schemes is higher than in global schemes: –Floating point programs contain more loop-control branches which exhibit periodic branch behavior. –This periodic branch behavior is better retained in multiple BHRs, as a result performance is better in per-address history schemes. Notes On Per-Address Schemes Performance

25. EECC722 - Shaaban #25 lec # 6 Fall 2005 9-26-2005 Variations of per-set history Two- Level Adaptive Branch Prediction BP-1 p = 2 m SPHTs b = 2 a PPHTs s = 2 i BHRs Here i = 2 s = 2 2 = 4 sets in level 1 Set Index bits

26. EECC722 - Shaaban #26 lec # 6 Fall 2005 9-26-2005 Performance of Per-set history schemes with different branch history lengths (2-18 bits) BP-1 9 SPEC89 benchmarks These curves are cut of when the implementation cost exceeds 512K bits. First level: 4 sets of BHRs (set index adress bits 10-11, same block of 1K address) SAg int SAg fp (SAp) (SAs) (SAg) (SAp) (SAs) (SAg)

27. EECC722 - Shaaban #27 lec # 6 Fall 2005 9-26-2005 Performance of Per-set history schemes with different number of pattern history tables (1-1024 PHTs) BP-1 9 SPEC89 benchmarks These curves are cut of when the implementation cost exceeds 512K bits. SAg 1 PHT i=2 set index bits 10, 11 used as branch set index to divide branches into s =4 sets in first level

28. EECC722 - Shaaban #28 lec # 6 Fall 2005 9-26-2005 Most Cost-effective Per Set: 8K: SAs(6, 4x16) 128K: SAs (9, 4x32)

29. EECC722 - Shaaban #29 lec # 6 Fall 2005 9-26-2005 Increasing the BHR length improves integer prediction accuracy significantly (similar to global schemes). The performance of per-set schemes is less sensitive to the increase of number of PHTs than in global schemes but more sensitive than per-address schemes. Floating point performance does not improve significantly when 4x16 or 4x256 PHTs are used. This is similar to per- address schemes. –This is due to partitioning the branches into sets (4 sets in this study using 4 BHTs) according to their addresses. Notes On Per-Set Schemes Performance/Cost

30. EECC722 - Shaaban #30 lec # 6 Fall 2005 9-26-2005 Comparison of the most effective configuration of each class of Two-Level Adaptive Branch Prediction with an implementation cost of 8K bits BP-1 Global: GAs (7, 32) Per-address: PAs (6, 16) Per-set: SAs (6, 4x16)

31. EECC722 - Shaaban #31 lec # 6 Fall 2005 9-26-2005 Comparison of the most effective configuration of each class of Two-Level Adaptive Branch Prediction with an implementation cost of 128K bits BP-1 Global: GAs (11, 32) Per-address: PAs (8, 256) Per-set: SAs (9, 4x32)

32. EECC722 - Shaaban #32 lec # 6 Fall 2005 9-26-2005 Performance Summary of Basic Two-level Branch Prediction Schemes Global history schemes perform better than other schemes on integer programs but require higher implementation costs to be effective overall –Integer programs contain many if-then-else branches which are effectively predicted by global schemes due to their better correlation with other previous branches –Global schemes require long BHR and or many PHTs to reduce interference for effective overall performance. Per-Address history schemes perform better than other schemes on floating point programs and require lower implementation costs to be effective. –Floating point programs contain many frequently-executed loop-control branches with periodic behavior which is better retained with a per-address BHT. –Pattern history of different branches interfere less than in global schemes, hence fewer PHTs are needed. Per-set schemes have integer performance similar to global schemes. They also have floating point performance similar to per-address schemes –To be effective, per-set schemes have even higher implementation costs due to the separate PHTs for each set (4 sets in this study).

33. EECC722 - Shaaban #33 lec # 6 Fall 2005 9-26-2005 Path-Based Prediction Ravi Nair proposed (1995) to use the path leading to a conditional branch rather than the branch history in the first level to index the PHT. The global history register of a GAp scheme is replaced by by a Path History Register, which encodes the addresses of the targets of the preceding p branches. The path history register could be organized as a g bit shift register which encodes q bits of the last p target addresses, where g = p x q. The hardware cost of such a mechanism is similar to that of a GAp scheme. If b branches are kept in the prediction data structure the cost is g + b x 2 x 2 g. The performance of this mechanism is similar or better than a comparable branch history schemes for the case of no context switching (worse with context switching). For the case that there is context switching, that is, if the processor switches between multiple processes running on the system, Nair proposes flushing the prediction data structures at regular intervals to improve accuracy. In such a scenario the mechanism performs slightly better than a comparable GAp predictor.

34. EECC722 - Shaaban #34 lec # 6 Fall 2005 9-26-2005 A Path-based Prediction Mechanism g = p x q PHR p branch targets q bits each a bits

35. EECC722 - Shaaban #35 lec # 6 Fall 2005 9-26-2005 Hybrid or Combined Predictors Hybrid predictors are simply combinations of other branch prediction mechanisms. This approach takes into account that different mechanisms may perform best for different branch scenarios (e.g. integer code vs floating point code). McFarling (1993) presented a number of different combinations of two branch prediction mechanisms. He proposed to use an additional 2-bit counter array which serves to select the appropriate predictor. One predictor is chosen for the higher two counts, the second one for the lower two counts. If the first predictor is wrong and the second one is right the counter is decremented, if the first one is right and the second one is wrong, the counter is incremented. No changes are carried out if both predictors are correct or wrong. BP-2

36. EECC722 - Shaaban #36 lec # 6 Fall 2005 9-26-2005 A Generic Hybrid Predictor BP-2 n is usually = 2

37. EECC722 - Shaaban #37 lec # 6 Fall 2005 9-26-2005 MCFarling’s Combined Predictor Structure The combined predictor contains an additional counter array with 2-bit up/down saturating counters. Which serves to select the best predictor to use. Each counter keeps track of which predictor is more accurate for the branches that share that counter. Specifically, using the notation P1c and P2c to denote whether predictors P1 and P2 are correct respectively, the counter is incremented or decremented by P1c-P2c as shown below. BP-2 X 11 10 01 00 Use P1 Use P2

38. EECC722 - Shaaban #38 lec # 6 Fall 2005 9-26-2005 MCFarling’s Combined Predictor Performance by Benchmark BP-2

39. EECC722 - Shaaban #39 lec # 6 Fall 2005 9-26-2005 MCFarling’s Combined Predictor Performance by Size BP-2 (Gap) (GAp) One Level

40. EECC722 - Shaaban #40 lec # 6 Fall 2005 9-26-2005 The Alpha 21264 Branch Prediction The Alpha 21264 uses a two-level adaptive hybrid method combining two algorithms (a global history and a per-branch history scheme) and chooses the best according to the type of branch instruction encountered The prediction table is associated with the lines of the instruction cache. An I-cache line contains 4 instructions along with a next line and a set predictor. If an I-cache line is fetched that contains a branch the next line will be fetched according to the line and set predictor. For lines containing no branches or unpredicted branches the next line predictor point simply to the next sequential cache line. This algorithm results in zero delay for correct predicted branches but wastes I-cache slots if the branch instruction is not in the last slot of the cache line or the target instruction is not in the first slot. The misprediction penalty for the alpha is 11 cycles on average and not less than 7 cycles. The resulting prediction accuracy is about 95% very good. Supports up to 6 branches in flight and employs a 32-entry return address stack for subroutines.

41. EECC722 - Shaaban #41 lec # 6 Fall 2005 9-26-2005 The Basic Alpha 21264 Pipeline

42. EECC722 - Shaaban #42 lec # 6 Fall 2005 9-26-2005 Alpha 21264 Branch Prediction

43. EECC722 - Shaaban #43 lec # 6 Fall 2005 9-26-2005 Sample Processor Branch Prediction Comparison Processor Released Accuracy Prediction Mechanism Cyrix 6x86 early '96 ca. 85% BHT associated with BTB Cyrix 6x86MX May '97 ca. 90% BHT associated with BTB AMD K5 mid '94 80% BHT associated with I-cache AMD K6 early '97 95% 2-level adaptive associated with BTIC and ALU Intel Pentium late '93 78% BHT associated with BTB Intel P6 mid '96 90% 2 level adaptive with BTB PowerPC750 mid '97 90% BHT associated with BTIC MC68060 mid '94 90% BHT associated with BTIC DEC Alpha early '97 95% Hybrid 2-level adaptive associated with I-cache HP PA8000 early '96 80% BHT associated with BTB SUN UltraSparc mid '95 88%int BHT associated with I-cache 94%FP

44. EECC722 - Shaaban #44 lec # 6 Fall 2005 9-26-2005 Branch Prediction Aliasing/Intereference Aliasing occurs when different branches point to the same prediction bits this leads to interference with current branches prediction by other branches. If the branch prediction mechanism is a one-level mechanism, k lower bits of the branch address are used to index the table. Two branches with the same lower k bits point to the same prediction bits. Similarly, in a PAg two-level scheme, the pattern history is indexed by the contents of history registers. If two branches have the same k history bits they will point to the same predictor entry in PHT. Three different cases of aliasing: –Constructive aliasing improves the prediction accuracy, –Destructive aliasing decreases the prediction accuracy, and –Harmless aliasing does not change the prediction accuracy. An alternative definition of aliasing applies the "three-C" model of cache misses to aliasing. Where aliasing is classified as three cases: –Compulsory aliasing occurs when a branch substream is encountered for the first time. – Capacity aliasing is due to the programs working set being too large to fit into the prediction data structures. Increasing the size of the data structures can reduce this effect. –Conflict aliasing occurs when two concurrently-active branch substreams map to the same predictor-table entry. Aliasing Interference

45. EECC722 - Shaaban #45 lec # 6 Fall 2005 9-26-2005 Aliasing /Interference in a Two-level Predictor Index to second level PHT formed from branch address, branch history BP-5 Branches A, B have the same index Aliasing Interference same index (aliasing) i.e Interference

46. EECC722 - Shaaban #46 lec # 6 Fall 2005 9-26-2005 Interference Reduction Prediction Schemes Gshare, 1993. (BP-2) The Filter Mechanism, 1996. (BP-3) The Agree Predictor, 1997. (BP-5) The Bi-Mode Predictor, 1997. (BP-6) The Skewed Branch Predictor, 1997. (BP-7) YAGS (Yet Another Global Scheme), 1998. (BP-10) BP-10 also covers a good overview of BP 5, 6, 7

47. EECC722 - Shaaban #47 lec # 6 Fall 2005 9-26-2005 MCFarling's gshare Predictor McFarling notes (1993) that using global history information might be less efficient than simply using the address of the branch instruction, especially for small predictors. He suggests using both global history and branch address by hashing them together. He proposes using the XOR of global branch history and branch address since he expects that this value has more information than either one of its components. This gives a more uniform distribution and usage of the PHT entries reducing conflict interference. The result is that this mechanism outperforms a GAp scheme by a small margin. This mechanism seems to use substantially less hardware, since both branch and pattern history are kept globally. The hardware cost for k history bits is the same as GAg: k + 2 x 2 k bits, neglecting costs for logic. BP-2 gshare = global history with index sharing

48. EECC722 - Shaaban #48 lec # 6 Fall 2005 9-26-2005 gshare Predictor Branch and pattern history are kept globally. History and branch address are XORed and the result is used to index the PHT. First Level Second Level BP-2 BHR PHT XOR

49. EECC722 - Shaaban #49 lec # 6 Fall 2005 9-26-2005 gshare Performance BP-2 GAg PAg SPEC89 used

50. EECC722 - Shaaban #50 lec # 6 Fall 2005 9-26-2005 The Agree Predictor The Agree Predictor is a scheme that tries to deal with the problem of aliasing, proposed by Sprangle, Chappell, Alsup and Patt They distinguish three approaches to counteract the interference problem: –Increasing predictor size to cause conflicting branches to map to different table locations. – Selecting a history table indexing function that distributes history states evenly among available counters. –Separating different classes of branches so that they do not use the same prediction scheme. The Agree predictor converts negative interference into positive or neutral interference by attaching a biasing bit to each branch, for instance in the BTB or instruction cache, which predicts the most likely outcome of the branch. The 2-bit counters of the branch prediction now predict whether or not the biasing bit is correct or not. The counters are updated after the branch has been resolved, agreement with the biasing bit leads to incrementing the counters, if there is no agreement the counter is decremented. The biasing bit could be determined by the direction the branch takes at the first occurrence, or by the direction most often taken by the branch. BP-5 BP-10

51. EECC722 - Shaaban #51 lec # 6 Fall 2005 9-26-2005 The Agree Predictor The hardware cost of this mechanism is that of the two-level adaptive mechanism it is based on, plus one bit per BTB entry or entry in the instruction cache. Simulations show that this scheme outperforms gshare, especially for smaller predictor sizes, because there is more contention than in bigger predictors. BP-5 BTB XNOR Prediction 0 0 1 0 1 0 1 0 0 1 1 1 Disagree Agree 1Agree 0 Disagree

52. EECC722 - Shaaban #52 lec # 6 Fall 2005 9-26-2005 Agree Predictor Operation BP-5 E.g. XOR as in gshare The 2-bit counters in PHT of the branch prediction predict whether or not the biasing bit is correct or not. 1 = agree 0 = disagree Offsets global predictors bad loop branch performance When branch outcome: Agrees with bias bit = increment Disagrees with bias bit = Decrement 0 0 1 0 1 0 1 0 0 1 1 1 Disagree Agree

53. EECC722 - Shaaban #53 lec # 6 Fall 2005 9-26-2005 Agree Predictor Performance vs. gshare Prediction accuracy versus predictor size for gcc. (gshare) BP-5 Small improvement for large PHT

54. EECC722 - Shaaban #54 lec # 6 Fall 2005 9-26-2005 Agree Predictor: Agree Predictor: Improvement in Accuracy from 1K to 64K Entry PHT gshare BP-5 Agree predictor has less interference than gshare  Thus its performance less sensitive to size of PHT

55. EECC722 - Shaaban #55 lec # 6 Fall 2005 9-26-2005 Agree Predictor: Branch Biasing Bit Selection The biasing bit could be determined by: –The direction the branch takes at the first occurrence “first time”, –Or by the direction most often taken by the branch “Most Often”. Performance results that the “first time” biasing bit selection method is comparable to “Most Often” selection method. BP-5 Less Diff.

56. EECC722 - Shaaban #56 lec # 6 Fall 2005 9-26-2005 The Bi-Mode Predictor The bi-mode predictor tries to replace destructive aliasing with neutral aliasing It splits the PHT table into even parts. One of the parts is the choice PHT, which is just a bimodal predictor with a slight change in the updating procedure. The other two parts are direction PHTs; one is a “taken” direction PHT and the other is a “not taken” direction PHT. The direction PHTs are indexed by the branch address XORed with the global history (similar to gshare). When a branch is present, its address points to the choice PHT entry which in turn chooses between the “taken” direction PHT and the “not taken” direction PHT. The prediction of the direction PHT chosen by the choice PHT serves as the prediction. Only the direction PHT chosen by the choice PHT is updated. The choice PHT is normally updated too, but not if it gives a prediction contradicting the branch outcome and the direction PHT chosen gives the correct prediction. As a result of this scheme, branches which are biased to be taken will have their predictions in the “taken” direction PHT, and branches which are biased not to be taken will have their predictions in the “not taken” direction PHT. So at any given time most of the information stored in the “taken” direction PHT entries is “taken” and any aliasing is more likely not to be destructive. In contrast to the agree predictor, if the bias is incorrectly chosen the first time the branch is introduced to the BTB, it is not bound to stay that way while the branch is in the BTB and as a result pollute the direction PHTs. However, it does not solve the aliasing problem between instances of a branch which do not agree with the bias and instances which do. BP-6 BP-10

57. EECC722 - Shaaban #57 lec # 6 Fall 2005 9-26-2005 The Bi-Mode Predictor BP-6 BP-10 BHR (First Level) X Choice PHT dynamically determines branch bias Second Level: Taken Direction PHT Not Taken Direction PHT

58. EECC722 - Shaaban #58 lec # 6 Fall 2005 9-26-2005 The Skewed Branch Predictor The skewed branch predictor is based on the observation that most aliasing occurs not because the size of the PHT is too small, but because of a lack of associativity in the PHT (the major contributor to aliasing is conflict aliasing and not capacity aliasing). The best way to deal with conflict aliasing is to make the PHT set-associative, but this requires tags and is not cost-effective. Instead, the skewed predictor emulates associativity using a special skewing function. The skewed branch predictor splits the PHT into three even banks and hashes each index to a 2-bit saturating counter in each bank using a unique hashing function per bank (f1, f2 and f3). The prediction is made according to a majority vote among the three banks. If the prediction is wrong all three banks are updated. If the prediction is correct, only the banks that made a correct prediction will be updated (partial updating). The skewing function should have inter-bank dispersion. This is needed to make sure that if a branch is aliased in one bank it will not be aliased in the other two banks, so the majority vote will produce an unaliased prediction. The reasoning behind partial updating is that if a bank gives a misprediction while the other two give correct predictions, the bank with the misprediction probably holds information which belongs to a different branch. In order to maintain the accuracy of the other branch, this bank is not updated. The skewed branch predictor tries to eliminate all aliasing instances and therefore all destructive aliasing. Unlike the other methods, it tries to eliminate destructive aliasing between branch instances which obey the bias and those which do not. Disadvantages: Redundancy of 1/3 - 2/3 of PHT size creates capacity aliasing. Slow to warm up on a context switch. BP-7 BP-10

59. EECC722 - Shaaban #59 lec # 6 Fall 2005 9-26-2005 The Skewed Branch Predictor BP-7 BP-10 Three different hashing functions f1, f2, f3 are used to index each PHT bank Second Level: Three PHT banks First Level

60. EECC722 - Shaaban #60 lec # 6 Fall 2005 9-26-2005 Yet Another Global Scheme (YAGS) As done in the agree and bi-mode, YAGS reduces aliasing by splitting the PHT into two branch streams corresponding to biases of “taken” and “not taken”. However, as in the skewed branch predictor, we do not want to neglect aliasing between biased branches and their instances which do not comply with the bias. The motivation behind YAGS is the observation that for each branch we need to store its bias and the instances when it does not agree with it. A bimodal predictor is used to store the bias, as the choice predictor does in the bi-mode scheme. All we need to store in the direction PHTs are the instances when the branch does not comply with its bias. To identify those instances in the direction PHTs small tags (6-8 bits) are added to each entry, referring to them (PHTs) now as direction caches. These tags store the least significant bits of the branch address and they virtually eliminate aliasing between two consecutive branches. When a branch occurs in the instruction stream, the choice PHT is accessed. If the choice PHT indicated “taken,” the “not taken” cache is accessed to check if it is a special case where the prediction does not agree with the bias. If there is a miss in the “not taken” cache, the choice PHT is used as a prediction. If there is a hit in the “not taken” cache it supplies the prediction. A similar set of actions is taken if the choice PHT indicates “not taken,” but this time the check is done in the “taken” cache. The choice PHT is addressed and updated as in the bi-mode choice PHT. The “not taken” cache is updated if a prediction from it was used. It is also updated if the choice PHT is indicating “taken” and the branch outcome was “not taken.” The same happens with the “taken” cache. Aliasing for instances of a branch which do not agree with the branch’s bias is reduced by making the direction caches set-associative: –LRU replacement is used policy with one exception: an entry in the “taken” cache which indicates “not taken” will be replaced first to avoid redundant information. If an entry in the “taken” cache indicates “not taken,” this information is already in the choice PHT and therefore is redundant and can be replaced. BP-10 PHT

61. EECC722 - Shaaban #61 lec # 6 Fall 2005 9-26-2005 YAGS BP-10 Direction Caches First Level Second Level

62. EECC722 - Shaaban #62 lec # 6 Fall 2005 9-26-2005 Performance of Four Interference Reduction Prediction Schemes YAGS6 (6 bits in the tags). BP-10 8 SPEC95 Benchmarks used

63. EECC722 - Shaaban #63 lec # 6 Fall 2005 9-26-2005 Performance of Four Interference Reduction Prediction Schemes Performance of Four Interference Reduction Prediction Schemes in The Presence of Context Switches. BP-10 8 SPEC95 Benchmarks used